La Forja: Adversarial Learning Studio for Math Reasoning

La Forja is described as an adversarial learning studio for high school and college mathematics, where students author math exam items rather than answer them. The system uses GPT-5.6 to challenge these items with reproducible counterexamples and evidence-based reviews, requiring students to repair and defend their work under pressure. It is built by one person (Manuel Peña Zuñiga) using technologies including Codex, GPT-5.6, JavaScript, Next.js, OpenAI, Prisma, React, and TypeScript.

The project is self-reported as a tool for fostering "builder thinking" in mathematics education, aiming to counter the trend of students using AI as an answer machine instead of developing reasoning skills. It positions itself as a mechanism to recreate peer-review dynamics from past educational communities like fmat.cl, where students defended solutions against scrutiny.

What changed: The description indicates a shift from traditional question-answering to item-authoring and adversarial review, with AI playing the role of reviewer rather than answer-provider. This is presented as an intentional pedagogical move to promote deeper understanding.

Single most important open question: Is there evidence that this model can scale beyond a single developer's prototype, or that it has traction in real-world educational settings?

The description states that La Forja is an adversarial learning studio for high school and college mathematics. Students do not answer questions; they author math exam items — including stems, alternatives, and rationales for distractors.

A reviewer (powered by GPT-5.6) attacks these items with:

• Three reviewers with different evidence contracts
• Checks for ambiguity, disciplinary errors, and weak distractors
• Deterministic probes for superficial answer cues

Every accepted finding must carry evidence: two conflicting interpretations that produce different answers, a cited source with the exact passage, or a reproducible heuristic.

The AI does not generate items or provide canonical solutions. Instead, it challenges the student’s item, which they then repair and defend in a short written viva scored on an explicit rubric.

Only after surviving this process is the item published with a "passport": provenance, challenges, revisions, and rubric results — all auditable.

The system implements:

• A versioned state machine
• Evidence contracts for reviewers
• Evaluation harness
• Deployment via Codex
• Deterministic checks that can never regress

It also includes:

• A labeled smoke set of sixteen original items with seeded flaws
• Comparison between a general reviewer and specialized gauntlet across repeated runs

Inference: The system appears to be a prototype built for demonstration purposes, not production-ready software.

The author claims that La Forja flips the direction of interaction in education. Rather than students using AI as an answer machine, it encourages them to build, argue, and defend ideas under pressure.

Key positioning elements:

• AI should not just speed up learning but structure productive struggle.
• The goal is to democratize not only access to teaching tools, but also the development of reasoning skills.
• Students must master content, anticipate how others go wrong, and design distractors around real misconceptions — this is described as "builder thinking."

The author notes that early drafts overclaimed:

• Guaranteed quality
• Synthetic psychometrics
• AI that never explains

These claims were cut after adversarial review.

Inference: The positioning has evolved from a broad vision to a more focused, defensible model centered on adversarial item creation and peer-like critique.

The description states that La Forja targets:

• High school and college students
• Educators preparing students for university entrance exams (specifically the PAES in Chile)

It is framed as a tool to help students develop reasoning skills, particularly in mathematics, where they must not only get the right answer but also defend it under challenge.

The system is described as exam-agnostic, suggesting potential applicability beyond Chile’s PAES.

Inference: The primary customer segment appears to be educators and learners in STEM-focused academic environments seeking deeper conceptual understanding through adversarial learning.

There is no evidence of a business model or pricing structure in the provided description. The project is presented as a hackathon submission, not a commercial product.

The author mentions building with Codex and GPT-5.6, but does not describe monetization strategies, licensing models, or customer acquisition plans.

Inference: No commercial viability or revenue model has been described.

The system is built using:

• Codex
• GPT-5.6
• JavaScript, Next.js, OpenAI, Prisma, React, TypeScript

Architecture details include:

• Versioned state machine
• Explicit concurrent calls with timeouts
• JSON schema validation on every model output
• Deterministic solver for probability domain of demo
• Application-level state machine where published versions are immutable and every repair creates a new version that must re-run the full history of accepted checks

The team built:

• A labeled smoke set of sixteen original items with seeded flaws
• Compared a single general reviewer against the specialized gauntlet across repeated runs, reporting exact counts, false positives, latency, and cost

Inference: The system is technically sophisticated for a prototype, but lacks evidence of scalability or production deployment.

There is no evidence of traction, revenue, customers, or adoption beyond the author’s own account. The project is described as a hackathon submission and a personal endeavor by one developer.

The description makes no mention of:

• Users
• Customers
• Revenue streams
• Market validation
• Product-market fit

Inference: No measurable traction or maturity indicators are evident.

There is no evidence in the description of competitors or competitive landscape. The author does not reference other platforms, tools, or systems used for math education or adversarial learning.

The project is positioned as a novel approach to math item creation and review, but no comparison with existing tools is made.

Inference: No competitive context is provided; it's unclear whether similar systems exist or how this one differentiates.

• Single developer: The entire system was built by one person (Manuel Peña Zuñiga), raising questions about scalability and long-term maintenance.
• Unverified claims: Many of the project’s assertions — such as its pedagogical impact or AI performance — are self-reported without external validation.
• No commercialization path: No evidence of a business model, pricing, or go-to-market strategy.
• Limited scope: The system is described as exam-agnostic but has no demonstrated use outside of Chile’s PAES.
• Technical complexity vs. prototype status: While technically advanced for a demo, there is no indication of production readiness.

Inference: The project is highly speculative and lacks any signs of traction or commercial viability.

1. What specific pedagogical outcomes have you observed in test environments?
2. How do you plan to scale beyond a single developer?
3. Have you conducted any user studies or pilot programs with students or educators?
4. What is your roadmap for monetization and product development?
5. How do you ensure consistency and fairness in AI-generated reviews across different types of items?
6. Are there any known limitations or biases in the current implementation that could affect educational equity?